Fast multi-body simulations of robots controlled with error feedback

Roboticists modeling control of manipulator and legged robots often assume force/torque-based control using an open-loop model of voltage, hydraulic pressure, or pneumatic pressure to actuator torques. This paper shows that such force/torque-based models can lead to stiff differential equations, which are computationally inefficient to solve: relatively small integration steps are necessary to ensure simulation stability. We have investigated two approaches that appear to mitigate this problem: incorporating transmission modeling (applicable only to electromagnetic actuators at present) and applying inverse dynamics control. Both of these approaches require increased computation per integration step, but this work will demonstrate that the larger integration steps can still yield considerably higher simulation throughput. This paper will also identify research challenges with using inverse dynamics control within multi-rigid body simulations, where the simulation is integrated forward subject to contact and inverse dynamics constraints. In particular, we examine a state of the art approach concerning inverse dynamics control, and we identify theoretical and practical algorithmic challenges. Experimental virtual robot platforms include a UR10 arm with attached prismatic manipulator and a locomoting quadrupedal robot.